Edge-Guided Remote-Sensing Image Compression

Using high-fidelity image compression makes it possible to transmit remote-sensing images in real-time. Nevertheless, existing lossy remote-sensing image compression (RSIC) methods have some inherent potential issues, including blocking and blurring effects, which are particularly problematic in low-compression-ratio (CR) settings. Although numerous methods have been studied to address the aforementioned issue, the majority of them exploit the prior of local smoothness in images, which usually induces the over-smoothing of regions with noticeable structure (i.e., edges and textures). During this task, we developed an innovative end-to-end framework that enables high-fidelity RSIC while retaining sharp edge and texture information. Initially, we put forth an edge-guided adversarial network (EGA-Net) for simultaneously restoring edge structures and generating texture details. Second, we impose an edge fidelity constraint to direct our network to optimize image content and structural information jointly. In addition, to facilitate this task, we have constructed a large-scale RSIC dataset named NWPU-RS-Compression (NWPU-RSC). This dataset contains over 300000 images of 30 categories, all with a fixed resolution of 600 × 600. Finally, a new quantitative metric for full reference image quality that takes into account signal statistics and the characteristics of the human visual system (HVS) has been developed, which helps evaluate reconstructed remote-sensing images more objectively and accurately. Experimental evidence has demonstrated that the EGA-Net surpasses several representative compression approaches regarding quality metrics on the NWPU-RSC, AID, and ISPR Vaihingen datasets. Code, dataset, and more experimental results can be accessed at https: //github.com/Chenxi1510/Remote-sensing-Image-Compression.